Fast and flexible sequence induction in spiking neural networks via rapid excitability changes

Rich Pang, Adrienne L. FairhalleLife 2019 · DOI

TL;DR

This paper proposes a novel mechanism for rapid cognitive flexibility in spiking neural networks, leveraging "long-term potentiation of intrinsic excitability" (LTP-IE) to quickly induce and replay sequences of neural activity. By combining fixed network structure with transient excitability changes, the model demonstrates how the brain might rapidly adapt to new information without slow, structural rewiring.

What Problem Does This Paper Address?

The brain's remarkable ability to adapt to new situations and learn new tasks quickly, known as cognitive flexibility, is a major puzzle. While we can rapidly form new associations (e.g., "raise left hand for word A, right hand for word B"), the underlying neural mechanisms are not well understood. Traditional models often rely on slow, gradual changes in synaptic weights, which are too slow for rapid adaptation. Furthermore, these slow changes risk "catastrophic forgetting" of existing memories. This paper investigates how faster, more dynamic mechanisms, specifically changes in neuronal excitability, could enable rapid sequence induction and flexible information processing in neural networks.

Methods

The authors developed a computational model of spiking neurons (leaky integrate-and-fire) to explore the role of LTP-IE. Their model incorporates three key components:

  1. Fixed Recurrent Connectivity: The underlying network structure of excitatory and inhibitory neurons remains constant.
  2. Stochastic Gating Inputs: Random, background inputs (G) are applied to the network, acting as a general "context" or "state" signal.
  3. Rapid LTP-IE: This is the core novel mechanism. When a neuron fires sufficiently rapidly (10-20 Hz) for a short period (1s) in response to sensory input (S), its intrinsic excitability is temporarily enhanced. This means its response to the gating inputs (G) is amplified, making it more likely to spike. This potentiation is modeled as a scaling factor ($\sigma$) applied to the gating input's effect.

They then used this model to simulate:

  • Sequence Replay: How previously experienced sensorimotor sequences (e.g., a trajectory through an environment) could be reactivated during rest.
  • Stimulus-Response Mapping: How LTP-IE could be used to create temporary associations between stimuli and responses in an untrained network.

Datasets

Not discussed in the paper. The study is a computational modeling work.

Key Insights and Ideas

  • Excitability as a Fast Switch: The paper highlights that changes in neuronal excitability, rather than just synaptic strength, can be a powerful and rapid mechanism for altering network dynamics. LTP-IE acts as a "tag" for recently active neurons, biasing them for future recruitment.
  • Fixed Structure, Dynamic Function: The model demonstrates that a stable, fixed network architecture can exhibit highly flexible and context-dependent behavior by modulating the excitability of its neurons. This offers a way to achieve rapid adaptation without the need for slow structural rewiring.
  • Replay as a Natural Outcome: The model shows that hippocampal-like replay (compressed, forward/reverse sequences) can emerge spontaneously from the interaction of LTP-IE-tagged neurons and fixed recurrent connectivity, even with random background inputs.
  • LTP-IE for Temporary Bindings: The research extends the concept to show how LTP-IE can create temporary stimulus-response mappings, suggesting a mechanism for rapid learning of new associations.

Key Results

  • Forward and Reverse Replay: The model successfully generated both forward and reverse replay of sensorimotor sequences, mirroring experimental observations in the hippocampus.
  • Robustness to Parameters: Sequence replay was observed across a wide range of network parameters, suggesting the mechanism is not overly sensitive to precise tuning.
  • Temporary Mappings: The model demonstrated the ability to encode and recall two distinct stimulus-response mappings (S1→M1, S2→M2 and S1→M2, S2→M1) by selectively applying LTP-IE to different sets of neurons. This was achieved even in a network without explicit spatial organization.
  • Speed and Order Reconstruction: The ordering of replayed sequences was determined by the fixed recurrent connectivity, not by the order in which LTP-IE was applied, suggesting a reconstruction of temporal information from network structure.

Limitations and Future Work

  • Decay of LTP-IE: The current model assumes LTP-IE is permanent within the simulation. The authors acknowledge that a decay mechanism for LTP-IE would be crucial for realistic memory consolidation and preventing all neurons from becoming equally excitable over time. They suggest that neuromodulatory inputs (like acetylcholine) could play a role in regulating LTP-IE decay.
  • Gating Input Structure: The gating inputs (G) were modeled as simple stochastic processes. Future work could explore how more complex, structured gating inputs (e.g., with history-dependent properties) might further enhance flexibility.
  • Biophysical Implementation: While the model captures the essence of LTP-IE, the precise biophysical implementation and its interaction with other plasticity mechanisms (like STDP) in real brains remain areas for further investigation.
  • Experimental Validation: The authors propose specific experimental predictions, such as blocking LTP-IE during exploration to prevent replay, which could be tested in future studies.
  • Generalization: The study focused on spatial sequences and simple mappings. Extending the model to more complex, non-spatial sequences (e.g., language) and exploring its applicability to other brain areas beyond the hippocampus are important future directions.

Summary updated

Abstract

Cognitive flexibility likely depends on modulation of the dynamics underlying how biological neural networks process information. While dynamics can be reshaped by gradually modifying connectivity, less is known about mechanisms operating on faster timescales. A compelling entrypoint to this problem is the observation that exploratory behaviors can rapidly cause selective hippocampal sequences to 'replay' during rest. Using a spiking network model, we asked whether simplified replay could arise from three biological components: fixed recurrent connectivity; stochastic 'gating' inputs; and rapid gating input scaling via long-term potentiation of intrinsic excitability (LTP-IE). Indeed, these enabled both forward and reverse replay of recent sensorimotor-evoked sequences, despite unchanged recurrent weights. LTP-IE 'tags' specific neurons with increased spiking probability under gating input, and ordering is reconstructed from recurrent connectivity. We further show how LTP-IE can implement temporary stimulus-response mappings. This elucidates a novel combination of mechanisms that might play a role in rapid cognitive flexibility.
BibTeX
@inproceedings{pang2019fast,
  author = {Rich Pang and Adrienne L. Fairhall},
  title = {{Fast and flexible sequence induction in spiking neural networks via rapid excitability changes}},
  booktitle = {eLife},
  year = {2019},
  doi = {10.7554/elife.44324},
}
Figure